Comparative Analysis of Gradient Boosting Algorithms for Credit Default Prediction in Microfinance

1. Introduction

Thin-file borrowers in microfinance settings often lack the credit bureau history that underpins conventional scorecards, prompting growing interest in alternative behavioural data sources and in gradient boosting methods that handle heterogeneous, sparse feature sets well.

2. Methodology

Three gradient boosting variants, XGBoost, LightGBM and CatBoost, were trained on 38,500 microloan records with 41 features spanning demographic, transactional and alternative behavioural data, and benchmarked against an L2-regularised logistic regression baseline using five-fold stratified cross-validation with AUC as the primary metric.

3. Results

CatBoost achieved the highest AUC at 0.879, followed by LightGBM at 0.874 and XGBoost at 0.869, with all three gradient boosting variants outperforming logistic regression at 0.792. CatBoost native handling of categorical features reduced feature-engineering and preprocessing time by an estimated 35 percent relative to the one-hot encoding pipeline required for the other models.

4. Conclusion

Gradient boosting methods, particularly CatBoost, offer meaningful accuracy gains over linear baselines for microfinance default prediction while reducing preprocessing overhead. Future work will examine fairness across demographic subgroups.

References

[1] Prokhorenkova L. et al., CatBoost: Unbiased boosting with categorical features, NeurIPS, 2018. [2] Ke G. et al., LightGBM, NeurIPS, 2017. [3] Bjorkegren D. and Grissen D., Behavior revealed in mobile phone usage predicts credit repayment, World Bank Economic Review, 2020.